The increased use of biometrics due to interest and acceptance have necessitated the development of accurate and efficient systems. This development is relevant to deal with large-scale systems whose usage have a trajectory to be increased further in the futures and are already in place in some capacity. These large-scale system consists of match candidates where there are a plethora of false match-candidates associated to each true match-candidate in the identification mode of biometric systems i.e. finding an identity within the biometric system and possibly who that identity belongs to. A good methodology to address the issue of accurately identifying true match-candidates is information fusion of biometric data from multiple biometric modalities. The issue that arises with the introduction of an information fusion in biometric systems is the huge workload. This motivates the idea of finding an intelligent way of applying information fusion in large-scale biometric systems that reduces the workload significantly while retaining the same (or even better) accuracy compared to an information fusion application on a full-scale system. The proposed intelligent way of applying information fusion in this project is a multi-stage multi-modal hierarchical k-stage system. This system pre-selects a shortlist of the best match-candidates, which are denoted by comparison scores from a given biometric modality, hierarchically in k-stages using different modalities at each stage. It is noted that the pre-selection is conducted on the shortlist denoted by the previous stage with the exception of the 1.level pre-selection which is performed on the full-scale system (i.e. the full-scale database or full-scale list) and the final level where the final selection (also called final decision) is made which is the level where the selection/ final match is conducted i.e. the decision determining if the claimed identity is within the system and who it belongs too. It is also stressed that a different biometric modality is used at each level. The assumption with this system is that it removes false match-candidates while retaining true match-candidates in concordance with significantly reducing the workload by removing false match-candidates from the system thus reducing the number of attempts (biometric identification attempts) to acquire the true match-candidate i.e. reduce the number of necessary biometric identification decisions.
The goal of the thesis is to investigate the effects of information fusion on largescale biometric system. The evaluation methodology was to establish a baseline consisting of individual biometric modalities and information fusion on the fullscale generated biometric system, and compare those baselines analytically to evaluations for configurations of the k-stage system that reduces the full-scale biometric system in size by multiple modalities over multiple levels. It is noted that a k-stage system configuration are a certain combination of modality orderings and pre-selection sizes. From the analytic comparison between baselines and k-stage configurations, it was possible to establish a model that will in terms of accuracy vs. efficiency denote the best k-stage system configurations.
The evaluations are based on ISO/IEC standard evaluation techniques such as Detection Error Trade-offs (DET) and Cumulative Match Characteristic (CMC) along with some common evaluation methodologies such as loss of genuines, possible number of biometric identification decisions and score distributions (for validation). Along with those evaluation techniques a workload reduction evaluation, whose metric for workload that has been proposed by Drozdowski et al., is utilized in this project to evaluate the workload reduction caused by the k-stage system applications. Subsequently, those evaluations techniques help illustrate the workload against accuracy trade-offs which represent the effects of efficiency vs. accuracy of the k-stage system compared to basic full-scale fusion techniques on large-scale biometric systems.